"""Resynthesis configuration — all fields are INITIAL VALUES, NOT CAPS. ``num_layers``, ``num_experts``, ``recursive_steps``, and domains are SEEDS that grow on observed demand. Termination is confidence-based (NeuralStopGate + utility/contradiction gates), never host-imposed caps. The ONE allowed backstop is the malfunction-guardrail (anti-infinite-loop), commented as guardrail-not-cap. Picking a bigger number is still a cap. """ from __future__ import annotations from dataclasses import dataclass # Resynthesis-owned vocabulary/feature parent (frozen, additive-only). # Its immutable inherited tensor geometry is hidden_size=4096, 32 layers, # linear_attention + full_attention hybrid, vocab=248320, native >4M context. # # AUTHORITY CONTRACT: these tensors are a small immutable lego in the larger # NoNE/RBO/Fabric graph. They may tokenize, project additive hidden state into # the inherited vocabulary, and seed feature/context tensors. Parent logits, # confidence, answer selection, and parent-relative retention are diagnostics # only and must never select, score, stop, retain, or veto an additive result. RESYNTHESIS_HIDDEN_SIZE = 4096 RESYNTHESIS_NUM_HIDDEN_LAYERS = 32 RESYNTHESIS_PROJECTION_VOCAB_SIZE = 248_320 # Long-context STACK+COMPOSE (compose, do not replace): # absolute KV order + Dual Chunk → parent RoPE indices + trained parent # YaRN live + exact online-softmax long pool over every hidden position. # Context-intent C, action A, and relational connectivity R: # Science attention composes C_i K_j, Q_i C_j, and R_i^Q R_j^K over the # preserved Q_i K_j path. C is model-owned hidden-plus-intent-plus-action # state: ``C = c_proj(hidden) + intent_c_proj(intent) + gate_a * action_c_proj(action)``. # R is conditioned by the NoNE router's selected expert-intent tensor. Exact # online-softmax tiling preserves every causal edge without a resident N x N # score table. Long pools compose intent and action as separate additive # ``Q·C`` and ``Q·A`` channels over the baseline ``Q·K`` path. These are # trained graph tensors, never host routing flags. # Long-pool C may be [batch, sequence] or [batch, sequence, hidden]; there is # no host cap on C/R width or sequence length (uncapped-policy: intentional). # Dual-Chunk keeps RoPE inside the pretrained band; do not YaRN-stretch. RESYNTHESIS_MAX_POSITION_EMBEDDINGS = 4_194_304 # Native attention must exceed the trained 4M floor. This is a proof/geometry # extent, not a floor+tile host cap: runtime admission observes absolute KV # extent while Dual Chunk preserves the parent's folded RoPE band and exact # KV-addressed history. RESYNTHESIS_NATIVE_CONTEXT_FLOOR = 4_194_304 NATIVE_ATTENTION_POSITION_APERTURE = 100_000_001 DUAL_CHUNK_PRETRAIN_LENGTH = 262_144 DUAL_CHUNK_LOCAL_SIZE = 8_192 # Dual Chunk keeps parent RoPE inside the pretrained band; do not YaRN-stretch. PRETRAINED_ROPE_BAND_TOKENS = DUAL_CHUNK_PRETRAIN_LENGTH # Execution tile width for parent/additive online-softmax pools. One tile equals # one Dual-Chunk successive window so prefill schedule and RoPE localization # stay coherent at 4M+ tokens. ``validate_native_context_admission`` validates # tensor shape/sign and observes full absolute KV extent; it must not silently # truncate or reject merely because the prompt passed the old floor+tile value. RESYNTHESIS_ONLINE_SOFTMAX_TILE_TOKENS = 8_192 # Native tiled prefill for the frozen parent backbone: long prompts are fed to # the complete parent graph in cache-preserving tiles so activations stay # bounded. Every token updates all 24 trained globally recurrent linear- # attention layers. Beyond the pretrained band, the eight full-attention # layers retain a bounded hot K/V window; no input token is dropped, rejected, # or truncated. One prefill tile equals one Dual-Chunk successive window and # one online-softmax pool window, keeping parent RoPE localization, recurrent # state, prefill scheduling, and the additive long pool coherent. RESYNTHESIS_NATIVE_PREFILL_TILE_TOKENS = 8_192 # Bounded activation geometry, not a context cap: every source token first # executes the complete parent and participates in an attended segment summary # before the additive Fabric/NoNE prefill traversal. RESYNTHESIS_NATIVE_PREFILL_SUMMARIES_PER_TILE = 4 RESYNTHESIS_NUM_ATTENTION_HEADS = 16 RESYNTHESIS_NUM_KEY_VALUE_HEADS = 4 RESYNTHESIS_INTERMEDIATE_SIZE = 12_288 RESYNTHESIS_HEAD_DIM = 256 RESYNTHESIS_FULL_ATTENTION_INTERVAL = 4 RESYNTHESIS_FULL_ATTENTION_LAYERS = ( RESYNTHESIS_NUM_HIDDEN_LAYERS // RESYNTHESIS_FULL_ATTENTION_INTERVAL ) RESYNTHESIS_LINEAR_ATTENTION_LAYERS = ( RESYNTHESIS_NUM_HIDDEN_LAYERS - RESYNTHESIS_FULL_ATTENTION_LAYERS ) # Long prompts retain the parent's globally recurrent linear-attention state at # every layer while bounding the resident full-attention K/V to one trained # Dual-Chunk local window. The pretrained-band boundary selects the storage # backend; it does not reject, truncate, route, or stop a prompt. RESYNTHESIS_NATIVE_HYBRID_ACTIVATION_TOKENS = DUAL_CHUNK_PRETRAIN_LENGTH RESYNTHESIS_NATIVE_HOT_KV_TOKENS = DUAL_CHUNK_LOCAL_SIZE # Resynthesis glyph-plane dimension. GLYPH_DIM = 168 # Public Release 188 resolves its immutable lexical parent and tokenizer # directly from repository-relative files. Historical lab manifests, receipts, # worktree paths, and source fingerprints are deliberately absent here. RESYNTHESIS_NATIVE_PARENT_GENERATION = "release_188" RESYNTHESIS_NATIVE_PARENT_ROOT = "weights/safetensors" RESYNTHESIS_NATIVE_PARENT_RUNTIME_ROOT = "" RESYNTHESIS_NATIVE_PARENT_MANIFEST_PATH = "" RESYNTHESIS_NATIVE_PARENT_MANIFEST_SHA256 = "" RESYNTHESIS_NATIVE_PARENT_MODEL_SHA256 = "" RESYNTHESIS_PARENT_CHECKPOINT_MANIFEST_PATH = "" RESYNTHESIS_PARENT_CHECKPOINT_MANIFEST_SHA256 = "" _RESYNTHESIS_PARENT_ARTIFACT_PACKAGE = "" # The public runtime uses the Transformers tokenizer files under ``tokenizer/``. # Fastokens remains an optional acceleration implementation rather than a # release-owned artifact or admission requirement. RESYNTHESIS_FASTOKENS_VERSION = "0.2.1" RESYNTHESIS_FASTOKENS_COMMIT = "" RESYNTHESIS_FASTOKENS_ROOT = "" RESYNTHESIS_FASTOKENS_SITE = "" RESYNTHESIS_FASTOKENS_RECEIPT_PATH = "" RESYNTHESIS_FASTOKENS_RECEIPT_SHA256 = "" RESYNTHESIS_FASTOKENS_SOURCE_ARCHIVE_SHA256 = "" RESYNTHESIS_FASTOKENS_WHEEL_PATH = "" RESYNTHESIS_FASTOKENS_WHEEL_SHA256 = "" RESYNTHESIS_PARENT_TOKENIZER_SHA256 = "" RESYNTHESIS_PARENT_TOKENIZER_VOCAB_SIZE = 248_077 @dataclass(frozen=True) class ResynthesisConfig: """Top-level config for the Resynthesis reasoning-science model. ``hidden_size`` names the frozen Resynthesis vocabulary/feature projection seam (4096), not the capacity of the complete additive model. Successor generations may add wider learned transfer, page, expert, and reasoning structure around that seam through explicit prefix-preserving migrations. Everything outside the immutable projection interface is an initial seed. """ # ── Frozen vocabulary/feature lego (never answer or learning authority) ── hidden_size: int = RESYNTHESIS_HIDDEN_SIZE # ``None`` starts the accepted additive generation at the complete current # feature seam. An explicit positive value is its initial materialized # knowledge-transfer rank, never a maximum. A retained successor may widen # this rank while keeping every accepted row/column byte-exact. knowledge_transfer_dim: int | None = None # Rank of the additive hidden-to-vocabulary residual pathway. ``None`` # derives the initial accepted rank from the model-owned knowledge-transfer # geometry, never from a fixed host constant. An explicit positive value is # likewise only an initial materialized rank: retained successors may widen # it through prefix-exact checkpoint migration, and no production path may # treat it as a maximum. logit_residual_rank: int | None = None base_model_dir: str = RESYNTHESIS_NATIVE_PARENT_ROOT native_parent_manifest_path: str = "" native_parent_manifest_sha256: str = "" # The graph parent predates the current rolling Resynthesis schema. These source # fields describe the historical reconstruction lineage for diagnostics; # they may change as the runtime improves and never veto load, resume, or # promotion. Exact model/checkpoint tensor identities remain authoritative. parent_runtime_source_dir: str = "" parent_source_fingerprint_path: str = "" parent_source_fingerprint_sha256: str = "" parent_source_id: str = "" parent_source_bundle_sha256: str = "" parent_source_recovery_manifest_path: str = "" parent_source_recovery_manifest_sha256: str = "" parent_rbo_outcome_source_path: str = "" parent_rbo_outcome_source_sha256: str = "" parent_integrated_source_sha256: str = "" parent_rbo_source_sha256: str = "" parent_none_fabric_source_sha256: str = "" parent_loader_source_sha256: str = "" # The 553-tensor capability payload is isolated from the authoritative # 1,043-tensor parent and fused by a separately trained Resynthesis surface. # The directory is audit provenance only: the executable compatibility # graph is pinned inside Resynthesis and never imports this mutable worktree. legacy_capability_source_dir: str = "" legacy_capability_source_sha256: str = "" base_weights: str = "model.safetensors" tokenizer_path: str = "tokenizer.json" config_path: str = "config.json" vocab_size: int = RESYNTHESIS_PROJECTION_VOCAB_SIZE # Exact lexical rows currently owned by the verified Fastokens BPE. The # frozen projection is padded to ``vocab_size`` for tensor geometry, so # these values are deliberately distinct. This is not a generation cap: # successor tokenizers may append rows while preserving every inherited # token ID, and the additive vocabulary bridge maps any remaining physical # projection rows through model-owned tensors before native selection. tokenizer_vocab_size: int = RESYNTHESIS_PARENT_TOKENIZER_VOCAB_SIZE # Immutable inherited token-ID prefix. Successor tokenizer artifacts may # append rows beyond both this prefix and the physical projection width, # but they may never renumber or replace these IDs. Keeping the prefix # separate from the active lexical size makes vocabulary growth explicit # without turning inherited projection geometry into a cap. tokenizer_prefix_vocab_size: int = ( RESYNTHESIS_PARENT_TOKENIZER_VOCAB_SIZE ) # Initial materialized rank of the learned projection-row-to-lexical-row # bridge. This is a growth seed, not a ceiling: successor generations may # widen it with prefix-exact checkpoint migration. The bridge prevents # projection-only coordinates from becoming undecodable token IDs while # retaining their trainable signal inside the Resynthesis graph. vocabulary_transfer_rank: int = 16 # ── Additive science MoE stack (SEEDS — grow on demand, no ceiling) ── num_layers: int = 4 # INITIAL count — grows via growth tool num_experts: int = 8 # INITIAL count — grows via growth tool expert_hidden_size: int = 1024 memory_slots: int = 1 # minimal seed row; grows on demand attention_heads: int = 16 mhc_heads: int = 8 # Boundary provenance only. This does not route or stop the model; it binds # a cold-reloaded graph to a checkpointed geometry migration. geometry_migrated: bool = False recursive_steps: int = 0 # 0 = RBO confidence controls traversal (no cap) residual_init: float = 0.02 # This scale gates an additional *additive-hidden* low-rank update. It no # longer blends parent logits into the answer surface; the frozen lm_head is # used only as the shared vocabulary projection over trained additive state. logit_residual_init: float = -4.0 # Parent-relative KL/identity retention is forbidden. Retention compares # against the prior accepted additive generation at the explicit learning # boundary, never against the hollow frozen vocabulary parent. kl_anchor_weight: float = 0.0 kl_anchor_warmup_steps: int = 100 glyph_input_dim: int = GLYPH_DIM # ── Identity ── identity_name: str = "Resynthesis" identity_creator: str = "Namenotfound.ai by Wendell Adams" # ── Execution grounding (default ON) ── execution_grounding: bool = True # ── Uncapped policy marker ── uncapped_policy: str = "intentional" # ── Model artifacts root ── artifacts_dir: str = ".nnf-resynthesis/models" # ── Exact public-release relocation coordinates ── # These paths may move in an exported repository. Their corresponding # identities remain the immutable native-parent/Fastokens authority. native_parent_model_sha256: str = "" parent_checkpoint_manifest_path: str = "" parent_checkpoint_manifest_sha256: str = "" fastokens_receipt_path: str = "" fastokens_receipt_sha256: str = "" fastokens_site_path: str = "" fastokens_wheel_path: str = "" fastokens_wheel_sha256: str = "" fastokens_source_archive_sha256: str = ""